Building Gen AI for Humans
I was recently in a user feedback session for a Gen AI product that we’ve been building here at BCG X.
Client: We love that the product can answer all your questions, but we don’t know what questions we can ask.
We are now at an inflection point in the build of Gen AI products. Up till now, it’s been a race to get something, anything, into production that utilizes Gen AI. In the race to bring these game-changing solutions to life, it has been a whirlwind for engineers and product teams alike, to keep up with the lightning-fast growth and evolution of this technology, often with little space for strategic design thought.
Now however, the expectation goes beyond “build me something with Gen AI”, to “build me something that my staff/customers actually want to use, with Gen AI”.
I’ve spent a bit of time thinking about this and discussing with my design and engineering colleagues. Here are 5 things we can do to move beyond chat interfaces, to start to build more user friendly iteractions with Gen AI.
1 — Use familiar interfaces that users already know how to use
Imagine a time, not so long ago, when search engines were new and novel. I actually remember listening to a radio segment when I was still at Primary school, featuring this new website called Google! The radio host was explaining how you could input words into a box and Google could return you articles related to your search.
Fast forward a little to University, and I remember attending a library seminar where the speaker was explaining how to return better results on Google by including a combination of quotes, asterisks and ampersands to better explain to the search engine what it was you were looking for. Don’t believe me? Here’s an article from 2008, explaining what I mean! Up until then, it was rather hit and miss on the quality of search results. These days, you would never expect to scroll beyond page 2 of the results to find exactly what you need.
Point is, people had to learn how to best use Google.
Gen AI isn’t yet mainstream, not every user has much experience with it yet, so they might need some help understanding how to interact with it.
With this in mind, we can design better user experience for Gen AI.
The empty prompt box which can return infinite possibilities is both exciting and overwhelming. Offering them a safety blanket on UI they already understand how to use can be very comforting.
Consider what products your clients already use. For example, if they are already using Microsoft Teams, can you borrow some of the interactions utilized there — ones that they already understand how to use — to indicate similar interactions in your app? Better still, can you integrate your Gen AI functionality directly within an app that they are already using?
2 — Provide clear instructions and training wheels for seamless interaction
There’s no point throwing brilliant technology at a problem, if your intended user doesn’t understand how to use it.
This point actually came directly out of our team discussions, following client feedback: it’s crucial to ensure users understand how to effectively utilize the tool you build.
Our Gen AI product was capable of equipping users with the ability to generate graphs effortlessly. However, we soon discovered that many users struggled to articulate their desired graph prompts to the chat functionality. This led to frustration when the incorrect graphs were generated, as well as valuable time being wasted.
To address this roadblock, our team put on their thinking caps and brainstormed a user-friendly solution: the “prompt builder.” Nothing more complex than drop-down menus where users could select the type of graph they desired, choose specific data sets, apply necessary filters, and decide what should be represented on the x and y axes.
The moment we introduced this simpler approach, the positive feedback poured in! By providing clear guardrails and step-by-step guidance, users found it much simpler to build effective prompts, minimizing the chances of generating the wrong graph, and the frustration that comes with it.
Happy users are much more likely to continuously return to your product.
3 — Utilise visual aids for better communication
The selling point of products like ChatGPT, is that it’s very easy to write a prompt that will return a block of text. The next level is to spend time improving the prompt to ensure we return the most accurate and useful answer. But it can’t stop there. Faced with chunks and chunks of returned text, how likely are you to read it all? The key lies in ensuring that users can easily digest and comprehend the generated information.
Consider this: When faced with a wall of text, it’s natural to feel overwhelmed and tempted to skim through or even skip the information altogether. To combat this, we started returning tables and graphs which went some of the way to return answers that are accessible and useful to users.
But we could go even further. What if we return HTML, or even an API-style response that we can populate a dashboard design? It’s all possible with Gen AI, it just relies on some tactical design thought, to take the user experience from good to excellent. It’s one thing to generate accurate and useful information, but ensuring that it is also easily understood and actionable, takes your product to a new level of success.
4 — Simplicity and personalisation: A balancing act
Of course one of the big benefits of Gen AI is that the output is almost infinite, enabling us the opportunity to provide highly personalised experiences for our users. However, the seemingly limitless range of customization can present a unique challenge — not only for engineers striving to deliver on these customizations but also for users who may feel overwhelmed by the vast array of options they have yet to explore.
I was recently working on a budgeting app. Of course with something as deeply individual as personal finance, you want to provide a highly custom experience. However bombarding users with an overwhelming universe of choices right from the start could lead to confusion and frustration.
To address this, I adopted a two-fold approach. First I introduced the user to a simple experience — a generic dashboard with basic functionality. This served as a starting point, allowing users to become familiar with the app’s core features before venturing into the realm of personalization. Once users felt comfortable navigating the basic functionality, we introduced an option to engage in a personalized chat. This allowed for a gradual transition, empowering users to unlock more enhanced features, once they felt comfortable to interact with them.
5 — Keep the flow as simple and as obvious as possible
Gen AI is a wonderful tool for improving efficiency. But if the users gets lost in your processes, your product wont be delivering on this capability.
How annoying is it when Alexa responds, “I’m sorry, I don’t know that one.” We can start to alleviate some of this frustration by offering our user options to steer them back to the happy path.
“I’m sorry, I can’t answer that, but I think you are asking X and I can point you in the direction of Y.”
Users will make mistakes and may take a while to understand your app, particularly in this new realm of Generative AI. However if you can keep gracefully guiding them back to the happy path, they are more likely to stick with your product.
As we stand at the forefront of this transformative technology, we have the opportunity to redefine the way users interact with AI. By focusing on familiar interfaces, clear instructions, visual aids, and simple, intuitive flows, we can shape products that not only bridge the gap between users and Gen AI but also surpass their expectations.
